Felix Kjellberg has entered the AI hardware race with Ajax, a compact language model designed to run entirely on local devices via his self-hosted Odysseus workspace. The YouTuber’s project bypasses the massive cloud servers dominating the industry, relying instead on Alibaba’s Qwen3.5-9B foundation. This nine-billion-parameter architecture keeps data processing close to the user, aiming for privacy and speed rather than raw computational power.
The announcement carries a distinct political edge. Kjellberg alleges that OpenAI suspended his account twice during the development phase, with one email explicitly citing “distillation” as the violation. Distillation involves using the outputs of a larger model to train a smaller one—a standard practice in machine learning but strictly regulated by OpenAI’s terms of service. The tech giant has not publicly confirmed or denied these suspensions, leaving Kjellberg’s account as the primary source for the conflict.
Ajax is not a general-purpose frontier model. It targets specific, daily tasks: web searches, email management, and calendar scheduling. Kjellberg describes the system as an “always-on” personal assistant. To achieve this, he reportedly used the open-source Heretic tool to “ablate” the model’s refusal behaviors, creating a more responsive interface. However, safety experts note that these modifications have not undergone independent assessment. The project remains in a pre-release phase, with the website currently listing the model as “Coming soon” and lacking final hardware requirements or downloadable weights.
This move signals a broader shift in the AI landscape. Users are increasingly wary of sending routine data to external servers. By building a local-first assistant, Kjellberg taps into this demand for data sovereignty. Yet, the project’s viability depends on resolving the legal gray areas surrounding model distillation. If OpenAI enforces its terms strictly, Ajax could become a test case for whether open-source communities can legally leverage proprietary model outputs.
Kjellberg plans to conduct further reinforcement learning and quantization before a public release. For now, Ajax serves as a prototype for a more personal, less corporate-centric approach to AI. The next step will be whether the model performs reliably on consumer hardware without the backing of big tech infrastructure.